Buy NVIDIA HGX servers with NVIDIA H200 for AI, LLM, and HPC online

Our NVIDIA HGX servers with NVIDIA H200 Tensor Core GPUs are designed for AI, LLM, and HPC workloads that push traditional GPU servers to their limits. Instead of individual PCIe cards, the HGX platform uses a shared SXM baseboard for up to eight GPUs, currently based on the NVIDIA H200. The GPUs communicate directly via NVLink at up to 900 GB/s per GPU, without the limitations found in traditional PCIe connections.

Why an NVIDIA HGX H200 server is the right choice for you

The combination of the HGX platform and the H200 GPU covers three application areas: training and inference of complex AI models, processing large language models with high memory requirements, and memory-intensive HPC tasks ranging from simulations to genome sequencing. The H200 is based on the Hopper architecture and features 141 GB of HBM3e memory and 4.8 terabytes per second (TB/s) of bandwidth—nearly double the capacity and 1.4 times the bandwidth of the H100. This accelerates the training and inference of large deep learning and language models such as Llama 2, as well as the aforementioned HPC tasks, where data bottlenecks would otherwise limit throughput. Despite its high performance, energy efficiency remains better than that of the previous generation, which reduces the total cost of ownership (TCO) in the data center.

Optimal Hardware Configuration for Your HGX H200 Server

A high-performance GPU server needs more than just the graphics cards themselves: AMD EPYC 9005 or Intel Xeon Scalable v5 processors prepare the data for the GPUs and keep them constantly utilized, while generous ECC memory ensures data integrity during continuous operation. Fast NVMe SSDs for the operating system and cache, combined with high-capacity data storage, ensure optimal data flow when loading large training datasets. To connect multiple servers into a cluster, NVIDIA ConnectX-7 and ConnectX-8 SuperNICs provide InfiniBand or high-speed Ethernet connectivity at up to 800 Gb/s per port, while BlueField DPUs offload network, storage, and security tasks from the main CPUs. Redundant power supplies and high-performance fans ensure availability even under full load, optionally supplemented by Direct Liquid Cooling interfaces for maximum packing density; hot-swap components simplify maintenance.

Your Custom NVIDIA HGX H200 Server

Whether it’s a research institution, university, or enterprise: every component can be freely configured. You’ll find additional options in the Server Technology Highlights. If you’re still unsure whether an HGX server is the right choice, our Server Buying Guide will help you with specific recommendations. In the Server Configurator, you can configure your system according to your individual needs.